Re: [PATCH] combine: Fix simplify_comparison AND handling for WORD_REGISTER_OPERATIONS targets [PR109040]

[Jeff Law <jeffreyalaw@gmail.com>]

[-- Attachment #1: Type: text/plain, Size: 1400 bytes --]



On 4/10/23 01:10, Jakub Jelinek wrote:
> On Sat, Apr 08, 2023 at 06:25:32PM -0600, Jeff Law wrote:
>>
>>
>> On 4/6/23 08:21, Eric Botcazou wrote:
>>
>>>> So, perhaps just in the return op0; case add further code for
>>>> WORD_REGISTER_OPERATIONS and sub-word modes which will call nonzero_bits
>>>> again for the word mode and decide if it is still safe.
>>>
>>> Does it work to just replace mode by word_mode in the calls to nonzero_bits?
>> It helps marginally -- basically we defer mucking up the code a bit.  We
>> then hit this in simplify_and_const_int_1:
>>
>>
>>    /* See what bits may be nonzero in VAROP.  Unlike the general case of
>>       a call to nonzero_bits, here we don't care about bits outside
>>       MODE.  */
>>
>>    nonzero = nonzero_bits (varop, mode) & GET_MODE_MASK (mode);
>>
>> That just seems wrong for WORD_REGISTER_OPERATIONS targets.
>>
>>
>> Hacking both locations in a similar manner fixes the test.
> 
> If so, can you post that in patch form and can we go with that version
> plus the testcase (e.g. from the first patch I've posted where I've changed
> dse)?
So as I mentioned in IRC, I haven't really looked closely at 
simplify_and_const_int_1.  I don't have a high degree of confidence this 
patch is complete, even though it does fix the test for 108947 and 109040.

I did bootstrap on riscv, but not a regression test, that's spinning 
right now.

Jeff

[-- Attachment #2: P --]
[-- Type: text/plain, Size: 1419 bytes --]

diff --git a/gcc/combine.cc b/gcc/combine.cc
index 22bf8e1ec89..c41d8a09b3b 100644
--- a/gcc/combine.cc
+++ b/gcc/combine.cc
@@ -10055,9 +10055,10 @@ simplify_and_const_int_1 (scalar_int_mode mode, rtx varop,
 
   /* See what bits may be nonzero in VAROP.  Unlike the general case of
      a call to nonzero_bits, here we don't care about bits outside
-     MODE.  */
+     MODE unless WORD_REGISTER_OPERATIONS is true.  */
 
-  nonzero = nonzero_bits (varop, mode) & GET_MODE_MASK (mode);
+  enum machine_mode tmode = WORD_REGISTER_OPERATIONS ? word_mode : mode;
+  nonzero = nonzero_bits (varop, tmode) & GET_MODE_MASK (tmode);
 
   /* Turn off all bits in the constant that are known to already be zero.
      Thus, if the AND isn't needed at all, we will have CONSTOP == NONZERO_BITS
diff --git a/gcc/simplify-rtx.cc b/gcc/simplify-rtx.cc
index 3b33afa2461..5f6f70491d8 100644
--- a/gcc/simplify-rtx.cc
+++ b/gcc/simplify-rtx.cc
@@ -3752,7 +3752,10 @@ simplify_context::simplify_binary_operation_1 (rtx_code code,
 	return op0;
       if (HWI_COMPUTABLE_MODE_P (mode))
 	{
-	  HOST_WIDE_INT nzop0 = nonzero_bits (trueop0, mode);
+	  /* When WORD_REGISTER_OPERATIONS is true, we need to know the
+	     nonzero bits in WORD_MODE rather than MODE.  */
+	  HOST_WIDE_INT nzop0
+	    = nonzero_bits (trueop0, WORD_REGISTER_OPERATIONS ? word_mode : mode);
 	  HOST_WIDE_INT nzop1;
 	  if (CONST_INT_P (trueop1))
 	    {